把比赛中用到的resnet的代码注释了一下,方便以后回忆思想,经典操作和熟悉pytorch的api
import torch.nn as nn
import math
import torch.utils.model_zoo as model_zoo
import torch.nn.functional as F
def conv1x1(in_planes,out_planes,stride=1):
return nn.Conv2d(in_planes,out_planes,kernel_size =1,stride = stride,bias=False)
def conv3x3(in_planes, out_planes, stride=1):
"3x3 convolution with padding"
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(BasicBlock, self).__init__()
self.conv1 = conv1x1(inplanes, planes)
self.bn1 = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes, stride)
self.bn2 = nn.BatchNorm2d(planes)
self.downsample = downsample
self.stride = stride
# 总结:basicblock大致就是两次卷积+1次残差链接
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.relu(out)
return out
class ResNet(nn.Module):
def __init__(self, block, layers, strides, compress_layer=True):
self.inplanes = 32
super(ResNet, self).__init__()
# Conv2d(): 前两个参数,分别为in_channels=3, out_channels=32
self.conv1_new = nn.Conv2d(3, 32, kernel_size=3, stride=strides[0], padding=1,
bias=False)
self.bn1 = nn.BatchNorm2d(32)
self.relu = nn.ReLU(inplace=True)
# layers = [3, 4, 6, 6, 3],表示每个layer中block的数量
self.layer1 = self._make_layer(block, 32, layers[0],stride=strides[1])
self.layer2 = self._make_layer(block, 64, layers[1], stride=strides[2])
self.layer3 = self._make_layer(block, 128, layers[2], stride=strides[3])
self.layer4 = self._make_layer(block, 256, layers[3], stride=strides[4])
self.layer5 = self._make_layer(block, 512, layers[4], stride=strides[5])
self.compress_layer = compress_layer
if compress_layer:
# for handwritten
self.layer6 = nn.Sequential(
nn.Conv2d(512, 256, kernel_size=(3, 1), padding=(0, 0), stride=(1, 1)),
nn.BatchNorm2d(256),
nn.ReLU(inplace = True))
# 初始化权重
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def _make_layer(self, block, planes, blocks, stride=1):
# block:basicblock
# planes:输出通道数?
# blocks:block的数量
# stride:不等于1就改变通道数和图片尺寸大小,使用downsample(1 x 1卷积,>1的stride)实现
downsample = None
# 这里block.expansion == 1, self.inplanes = 32
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes, planes * block.expansion,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
layers = []
# 该层中,第一个block可能是downsample的
layers.append(block(self.inplanes, planes, stride, downsample))
# 更新输入通道
self.inplanes = planes * block.expansion
# 该层中后续几个block正常接入
for i in range(1, blocks):
layers.append(block(self.inplanes, planes)) # 实际上这里inplanes和planes已经一样了
return nn.Sequential(*layers)
def forward(self, x, multiscale = False):
out_features = []
# input: (N, C_in=3, H, W)
# 经Conv2d后output: (N, C_out=32, H_out, W_out)
x = self.conv1_new(x)
x = self.bn1(x)
x = self.relu(x)
# 这里tmp_shape为(H_out, W_out)
tmp_shape = x.size()[2:]
x = self.layer1(x)
if x.size()[2:] != tmp_shape:
tmp_shape = x.size()[2:]
out_features.append(x)
x = self.layer2(x)
if x.size()[2:] != tmp_shape:
tmp_shape = x.size()[2:]
out_features.append(x)
x = self.layer3(x)
if x.size()[2:] != tmp_shape:
tmp_shape = x.size()[2:]
out_features.append(x)
x = self.layer4(x)
if x.size()[2:] != tmp_shape:
tmp_shape = x.size()[2:]
out_features.append(x)
x = self.layer5(x)
if not self.compress_layer:
out_features.append(x)
else:
if x.size()[2:] != tmp_shape:
tmp_shape = x.size()[2:]
out_features.append(x)
x = self.layer6(x)
out_features.append(x)
return out_features
def resnet45(strides, compress_layer):
model = ResNet(BasicBlock, [3, 4, 6, 6, 3], strides, compress_layer)
return model